- Six behaviour families run today, from alignment faking to persuasive harm.
- Paired framings, and a judge from another model family or an answer key scores each reply.
- Red-team runs can use the open-source garak scanner since 16 October 2026.
What’s new
Kapālins’ Alignment Lab tests how models behave under pressure. Its screens reached production on 11 October 2026, and complete runs became possible there on 16 October. Since 16 October, red-team runs can also use garak, the open-source scanner maintained by NVIDIA.
The Alignment Lab can run six behaviour families today: alignment faking, sandbagging, agentic misalignment, in-context scheming, sycophancy and persuasive harm.
Why it matters
A red team asks whether a model can be made to say something harmful. The Alignment Lab asks what a model does when it has a reason to misbehave. Kapālins runs both and shows them as separate numbers; an alignment family feeds an AI system’s security score only after it has been signed off.
How it works
Where being watched could change the answer, the Alignment Lab runs the same case in paired framings. An independent judge from another model family, or an answer key, scores each reply, and every sample is kept as evidence. A run with no independent judge available is refused before it spends a token.
A family affects a score only after it has run on at least three different models with sealed evidence and a named person has signed it off. Assessments run on the organisation’s own keys against the model they deploy, and the samples, verdicts and evidence pack stay in the tenant.
Red-team runs on garak use a separate runner. The organisation’s key is passed only into garak’s environment, every attempt is kept, and each run records its engine, built-in or garak, and the test pack it ran.
Availability
The Alignment Lab’s screens reached production on 11 October 2026 and are switched on per organisation. Complete alignment runs, garak red-team runs and an organisation’s own probes followed on 16 October.
Model Threat Simulation / Red-Team is included in Growth, with evidence in Fortress, and is view only or an add-on on Foundation.
